Fix performance problems with autovacuum truncation in busy workloads.

In situations where there are over 8MB of empty pages at the end of
a table, the truncation work for trailing empty pages takes longer
than deadlock_timeout, and there is frequent access to the table by
processes other than autovacuum, there was a problem with the
autovacuum worker process being canceled by the deadlock checking
code. The truncation work done by autovacuum up that point was
lost, and the attempt tried again by a later autovacuum worker. The
attempts could continue indefinitely without making progress,
consuming resources and blocking other processes for up to
deadlock_timeout each time.

This patch has the autovacuum worker checking whether it is
blocking any other thread at 20ms intervals. If such a condition
develops, the autovacuum worker will persist the work it has done
so far, release its lock on the table, and sleep in 50ms intervals
for up to 5 seconds, hoping to be able to re-acquire the lock and
try again. If it is unable to get the lock in that time, it moves
on and a worker will try to continue later from the point this one
left off.

While this patch doesn't change the rules about when and what to
truncate, it does cause the truncation to occur sooner, with less
blocking, and with the consumption of fewer resources when there is
contention for the table's lock.

The only user-visible change other than improved performance is
that the table size during truncation may change incrementally
instead of just once.

This problem exists in all supported versions but is infrequently
reported, although some reports of performance problems when
autovacuum runs might be caused by this. Initial commit is just the
master branch, but this should probably be backpatched once the
build farm and general developer usage confirm that there are no
surprising effects.

Jan Wieck

1 files changed, 166 insertions, 64 deletions

diff --git a/src/backend/commands/vacuumlazy.c b/src/backend/commands/vacuumlazy.c
index 503684936b3..fc18b27427b 100644
--- a/src/backend/commands/vacuumlazy.c
+++ b/src/backend/commands/vacuumlazy.c
@@ -48,6 +48,7 @@
 #include "commands/vacuum.h"
 #include "miscadmin.h"
 #include "pgstat.h"
+#include "portability/instr_time.h"
 #include "postmaster/autovacuum.h"
 #include "storage/bufmgr.h"
 #include "storage/freespace.h"
@@ -70,6 +71,17 @@
 #define REL_TRUNCATE_FRACTION	16
 
 /*
+ * Timing parameters for truncate locking heuristics.
+ *
+ * These were not exposed as user tunable GUC values because it didn't seem
+ * that the potential for improvement was great enough to merit the cost of
+ * supporting them.
+ */
+#define AUTOVACUUM_TRUNCATE_LOCK_CHECK_INTERVAL		20	/* ms */
+#define AUTOVACUUM_TRUNCATE_LOCK_WAIT_INTERVAL		50	/* ms */
+#define AUTOVACUUM_TRUNCATE_LOCK_TIMEOUT			5000		/* ms */
+
+/*
  * Guesstimation of number of dead tuples per page.  This is used to
  * provide an upper limit to memory allocated when vacuuming small
  * tables.
@@ -103,6 +115,7 @@ typedef struct LVRelStats
 	ItemPointer dead_tuples;	/* array of ItemPointerData */
 	int			num_index_scans;
 	TransactionId latestRemovedXid;
+	bool		lock_waiter_detected;
 } LVRelStats;
 
 
@@ -193,6 +206,8 @@ lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
 	vacrelstats->old_rel_pages = onerel->rd_rel->relpages;
 	vacrelstats->old_rel_tuples = onerel->rd_rel->reltuples;
 	vacrelstats->num_index_scans = 0;
+	vacrelstats->pages_removed = 0;
+	vacrelstats->lock_waiter_detected = false;
 
 	/* Open all indexes of the relation */
 	vac_open_indexes(onerel, RowExclusiveLock, &nindexes, &Irel);
@@ -259,10 +274,17 @@ lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
 						vacrelstats->hasindex,
 						new_frozen_xid);
 
-	/* report results to the stats collector, too */
-	pgstat_report_vacuum(RelationGetRelid(onerel),
-						 onerel->rd_rel->relisshared,
-						 new_rel_tuples);
+	/*
+	 * Report results to the stats collector, too. An early terminated
+	 * lazy_truncate_heap attempt suppresses the message and also cancels the
+	 * execution of ANALYZE, if that was ordered.
+	 */
+	if (!vacrelstats->lock_waiter_detected)
+		pgstat_report_vacuum(RelationGetRelid(onerel),
+							 onerel->rd_rel->relisshared,
+							 new_rel_tuples);
+	else
+		vacstmt->options &= ~VACOPT_ANALYZE;
 
 	/* and log the action if appropriate */
 	if (IsAutoVacuumWorkerProcess() && Log_autovacuum_min_duration >= 0)
@@ -1257,80 +1279,124 @@ lazy_truncate_heap(Relation onerel, LVRelStats *vacrelstats)
 	BlockNumber old_rel_pages = vacrelstats->rel_pages;
 	BlockNumber new_rel_pages;
 	PGRUsage	ru0;
+	int			lock_retry;
 
 	pg_rusage_init(&ru0);
 
 	/*
-	 * We need full exclusive lock on the relation in order to do truncation.
-	 * If we can't get it, give up rather than waiting --- we don't want to
-	 * block other backends, and we don't want to deadlock (which is quite
-	 * possible considering we already hold a lower-grade lock).
-	 */
-	if (!ConditionalLockRelation(onerel, AccessExclusiveLock))
-		return;
-
-	/*
-	 * Now that we have exclusive lock, look to see if the rel has grown
-	 * whilst we were vacuuming with non-exclusive lock.  If so, give up; the
-	 * newly added pages presumably contain non-deletable tuples.
+	 * Loop until no more truncating can be done.
 	 */
-	new_rel_pages = RelationGetNumberOfBlocks(onerel);
-	if (new_rel_pages != old_rel_pages)
+	do
 	{
 		/*
-		 * Note: we intentionally don't update vacrelstats->rel_pages with the
-		 * new rel size here.  If we did, it would amount to assuming that the
-		 * new pages are empty, which is unlikely.	Leaving the numbers alone
-		 * amounts to assuming that the new pages have the same tuple density
-		 * as existing ones, which is less unlikely.
+		 * We need full exclusive lock on the relation in order to do
+		 * truncation. If we can't get it, give up rather than waiting --- we
+		 * don't want to block other backends, and we don't want to deadlock
+		 * (which is quite possible considering we already hold a lower-grade
+		 * lock).
 		 */
-		UnlockRelation(onerel, AccessExclusiveLock);
-		return;
-	}
+		vacrelstats->lock_waiter_detected = false;
+		lock_retry = 0;
+		while (true)
+		{
+			if (ConditionalLockRelation(onerel, AccessExclusiveLock))
+				break;
 
-	/*
-	 * Scan backwards from the end to verify that the end pages actually
-	 * contain no tuples.  This is *necessary*, not optional, because other
-	 * backends could have added tuples to these pages whilst we were
-	 * vacuuming.
-	 */
-	new_rel_pages = count_nondeletable_pages(onerel, vacrelstats);
+			/*
+			 * Check for interrupts while trying to (re-)acquire the exclusive
+			 * lock.
+			 */
+			CHECK_FOR_INTERRUPTS();
 
-	if (new_rel_pages >= old_rel_pages)
-	{
-		/* can't do anything after all */
-		UnlockRelation(onerel, AccessExclusiveLock);
-		return;
-	}
+			if (++lock_retry > (AUTOVACUUM_TRUNCATE_LOCK_TIMEOUT /
+								AUTOVACUUM_TRUNCATE_LOCK_WAIT_INTERVAL))
+			{
+				/*
+				 * We failed to establish the lock in the specified number of
+				 * retries. This means we give up truncating. Suppress the
+				 * ANALYZE step. Doing an ANALYZE at this point will reset the
+				 * dead_tuple_count in the stats collector, so we will not get
+				 * called by the autovacuum launcher again to do the truncate.
+				 */
+				vacrelstats->lock_waiter_detected = true;
+				ereport(LOG,
+						(errmsg("automatic vacuum of table \"%s.%s.%s\": "
+								"cannot (re)acquire exclusive "
+								"lock for truncate scan",
+								get_database_name(MyDatabaseId),
+							get_namespace_name(RelationGetNamespace(onerel)),
+								RelationGetRelationName(onerel))));
+				return;
+			}
 
-	/*
-	 * Okay to truncate.
-	 */
-	RelationTruncate(onerel, new_rel_pages);
+			pg_usleep(AUTOVACUUM_TRUNCATE_LOCK_WAIT_INTERVAL);
+		}
 
-	/*
-	 * We can release the exclusive lock as soon as we have truncated.	Other
-	 * backends can't safely access the relation until they have processed the
-	 * smgr invalidation that smgrtruncate sent out ... but that should happen
-	 * as part of standard invalidation processing once they acquire lock on
-	 * the relation.
-	 */
-	UnlockRelation(onerel, AccessExclusiveLock);
+		/*
+		 * Now that we have exclusive lock, look to see if the rel has grown
+		 * whilst we were vacuuming with non-exclusive lock.  If so, give up;
+		 * the newly added pages presumably contain non-deletable tuples.
+		 */
+		new_rel_pages = RelationGetNumberOfBlocks(onerel);
+		if (new_rel_pages != old_rel_pages)
+		{
+			/*
+			 * Note: we intentionally don't update vacrelstats->rel_pages with
+			 * the new rel size here.  If we did, it would amount to assuming
+			 * that the new pages are empty, which is unlikely. Leaving the
+			 * numbers alone amounts to assuming that the new pages have the
+			 * same tuple density as existing ones, which is less unlikely.
+			 */
+			UnlockRelation(onerel, AccessExclusiveLock);
+			return;
+		}
 
-	/*
-	 * Update statistics.  Here, it *is* correct to adjust rel_pages without
-	 * also touching reltuples, since the tuple count wasn't changed by the
-	 * truncation.
-	 */
-	vacrelstats->rel_pages = new_rel_pages;
-	vacrelstats->pages_removed = old_rel_pages - new_rel_pages;
+		/*
+		 * Scan backwards from the end to verify that the end pages actually
+		 * contain no tuples.  This is *necessary*, not optional, because
+		 * other backends could have added tuples to these pages whilst we
+		 * were vacuuming.
+		 */
+		new_rel_pages = count_nondeletable_pages(onerel, vacrelstats);
 
-	ereport(elevel,
-			(errmsg("\"%s\": truncated %u to %u pages",
-					RelationGetRelationName(onerel),
-					old_rel_pages, new_rel_pages),
-			 errdetail("%s.",
-					   pg_rusage_show(&ru0))));
+		if (new_rel_pages >= old_rel_pages)
+		{
+			/* can't do anything after all */
+			UnlockRelation(onerel, AccessExclusiveLock);
+			return;
+		}
+
+		/*
+		 * Okay to truncate.
+		 */
+		RelationTruncate(onerel, new_rel_pages);
+
+		/*
+		 * We can release the exclusive lock as soon as we have truncated.
+		 * Other backends can't safely access the relation until they have
+		 * processed the smgr invalidation that smgrtruncate sent out ... but
+		 * that should happen as part of standard invalidation processing once
+		 * they acquire lock on the relation.
+		 */
+		UnlockRelation(onerel, AccessExclusiveLock);
+
+		/*
+		 * Update statistics.  Here, it *is* correct to adjust rel_pages
+		 * without also touching reltuples, since the tuple count wasn't
+		 * changed by the truncation.
+		 */
+		vacrelstats->pages_removed += old_rel_pages - new_rel_pages;
+		vacrelstats->rel_pages = new_rel_pages;
+
+		ereport(elevel,
+				(errmsg("\"%s\": truncated %u to %u pages",
+						RelationGetRelationName(onerel),
+						old_rel_pages, new_rel_pages),
+				 errdetail("%s.",
+						   pg_rusage_show(&ru0))));
+		old_rel_pages = new_rel_pages;
+	} while (new_rel_pages > vacrelstats->nonempty_pages &&
+			 vacrelstats->lock_waiter_detected);
 }
 
 /*
@@ -1342,6 +1408,12 @@ static BlockNumber
 count_nondeletable_pages(Relation onerel, LVRelStats *vacrelstats)
 {
 	BlockNumber blkno;
+	instr_time	starttime;
+	instr_time	currenttime;
+	instr_time	elapsed;
+
+	/* Initialize the starttime if we check for conflicting lock requests */
+	INSTR_TIME_SET_CURRENT(starttime);
 
 	/* Strange coding of loop control is needed because blkno is unsigned */
 	blkno = vacrelstats->rel_pages;
@@ -1354,6 +1426,36 @@ count_nondeletable_pages(Relation onerel, LVRelStats *vacrelstats)
 		bool		hastup;
 
 		/*
+		 * Check if another process requests a lock on our relation. We are
+		 * holding an AccessExclusiveLock here, so they will be waiting. We
+		 * only do this in autovacuum_truncate_lock_check millisecond
+		 * intervals, and we only check if that interval has elapsed once
+		 * every 32 blocks to keep the number of system calls and actual
+		 * shared lock table lookups to a minimum.
+		 */
+		if ((blkno % 32) == 0)
+		{
+			INSTR_TIME_SET_CURRENT(currenttime);
+			elapsed = currenttime;
+			INSTR_TIME_SUBTRACT(elapsed, starttime);
+			if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000)
+				>= AUTOVACUUM_TRUNCATE_LOCK_CHECK_INTERVAL)
+			{
+				if (LockHasWaitersRelation(onerel, AccessExclusiveLock))
+				{
+					ereport(elevel,
+							(errmsg("\"%s\": suspending truncate "
+									"due to conflicting lock request",
+									RelationGetRelationName(onerel))));
+
+					vacrelstats->lock_waiter_detected = true;
+					return blkno;
+				}
+				starttime = currenttime;
+			}
+		}
+
+		/*
 		 * We don't insert a vacuum delay point here, because we have an
 		 * exclusive lock on the table which we want to hold for as short a
 		 * time as possible.  We still need to check for interrupts however.